I've been asked about my 2024 predictions in AI/Robotics 1) Humanoid companies bifurcate into industrial/structured and unstructured/human-centric based on hardware directions 2) We finally see generalization in RoboticsAI 3) Signs of embodied data significantly improving LLMs 🧵

Stiff humanoids (most of them) will encounter problems deploying safe manipulation and operating in the messy uncertainty of the world but find applications in structured environments where calibration and adaptation of the environment are feasible (factories/logistics)

Compliant/soft systems will leverage the advantage of safe manipulation and efficient operation in unstructured environments to gather more diverse data (consumer and services in addition to the industrial). The jury is still out on if this is required, but this is our bet at 1X

Manipulation has not yet seen generalization like Pix2Pix/LLM (RT-X while impressive was a negative result on increasing dexterity from non-robot data). Deployments at scale will enable generalization through large-scale immitation learning (For 1X, NEOs expansion into consumer)

Signs that LLMs greatly improve with the same embodied data (how actions changes the state of the world, modalities like egocentric touch/audio/vision). But no one is yet training foundation models on this combined data due to the large players staying clear of hardware

@BerntBornich What timeline are you thinking of for humanoid bots being say 25-50% of the workforce in factories?